Pseudo trajectories eliminating and pyramid clustering: Optimizing dense trajectories for action recognition

摘要

Motion path methods, especially for dense trajectories, have showed effectiveness for action recognition. To select the most representative trajectories of motion objects, a novel approach is presented to eliminate the pseudo trajectories generated by background disturbance, camera shake and failure tracking. Benefits in our method include: (i) Motion region boosting compensating object shift. (ii) Differential thresholding for rapid computation. (iii) Strengthening the independence of words in Bag of Features by pyramid clustering. Experimental results prove that our method selects dense trajectories belonging to objects effectively and improves the recognition accuracy.